iq_train_set / README.md
noshou's picture
Upload README.md with huggingface_hub
c055453 verified
|
Raw
History Blame Contribute Delete
10.8 kB
---
license: cc0-1.0
task_categories:
- other
tags:
- chemistry
- scattering
- molecular
- hdf5
pretty_name: IQ Train Set (L=50)
size_categories:
- 1M<n<10M
---
# I(q)@L=50 Database
## Parameters
| Parameter | Value |
| --------- | ---------------- |
| energy | 12 500 eV |
| qMin | 0 angstrom^-1 |
| qMax | 0.5 angstrom^-1 |
| step | 0.01 angstrom^-1 |
| Q | 51 points |
| lMax | 50 |
## Files
| File | Size | Description |
|------|------|-------------|
| `I(q)@L=50.h5` | ~66 GB | HDF5 database of I(q) curves and molecular data |
| `iq_train_set-ENCODING.sqlite3` | ~860 MB | Encoding index: maps every molecule to its atom count and VOCAB indices, so the data pipeline never needs to scan the 66 GB HDF5 file during training |
| `xyz_coordinate_files.7z` | ~6.5 GB | Source XYZ geometry files for all molecule groups (LZMA2, max compression). Only needed to re-run the build pipeline from scratch. |
## Retrieving the dataset
The HDF5 files are hosted on **[HuggingFace (noshou/iq_train_set)](https://huggingface.co/datasets/noshou/iq_train_set)** and **[Kaggle (noso0s0n/iql50)](https://www.kaggle.com/datasets/noso0s0n/iql50)**. The training code (ScatterNet model, preprocessing pipeline, baselines) lives in the **[noshou/APS360](https://github.com/noshou/APS360)** GitHub repository; the `Preprocess/` directory contains the encoding and data pipeline code.
Download both the HDF5 file and the encoding DB with the HuggingFace CLI (recommended - resumes interrupted downloads):
```bash
pip install huggingface_hub
hf download noshou/iq_train_set "I(q)@L=50.h5" "iq_train_set-ENCODING.sqlite3" \
--repo-type dataset --local-dir Preprocess/
```
Or in Python:
```python
from huggingface_hub import hf_hub_download
for filename in ["I(q)@L=50.h5", "iq_train_set-ENCODING.sqlite3"]:
hf_hub_download(
repo_id = "noshou/iq_train_set",
filename = filename,
repo_type = "dataset",
local_dir = "Preprocess/",
)
```
Both files are also available on the **[Kaggle dataset](https://www.kaggle.com/datasets/noso0s0n/iql50)** and are mounted directly as notebook inputs when using `kaggle_train.ipynb` / `kaggle_baselines.ipynb` -- no download step needed there.
Place the downloaded files at `Preprocess/I(q)@L=50.h5` and `Preprocess/iq_train_set-ENCODING.sqlite3` (the paths all pipeline scripts expect).
## Running training
### Local (CLI)
Edit `Train/train.yaml` to set paths, then:
```bash
python Train/train.py --config Train/train.yaml
```
Key paths in `train.yaml`:
```yaml
hdf5: Preprocess/I(q)@L=50.h5 # downloaded above
encodings_sqlite3_path: Preprocess/iq_train_set-ENCODING.sqlite3 # downloaded above
```
### Kaggle (notebook)
Open `Baselines/kaggle_baselines.ipynb`. Set `NOTEBOOK_NAME` to your Kaggle notebook slug at the top of the setup cell, and attach the [`noso0s0n/iql50`](https://www.kaggle.com/datasets/noso0s0n/iql50) dataset as a notebook input -- it provides both `I(q)@L=50.h5` and `iq_train_set-ENCODING.sqlite3` pre-mounted under `/kaggle/input/datasets/noso0s0n/iql50/`, no download or build step needed. The notebook clones the repo, installs dependencies, and runs all baselines.
---
Produced by `buildDB()` in `load_data.py`. The file is opened in append mode (`'a'`), so existing entries are skipped on resume.
## Root attributes
| Attribute | Type | Description |
| --------- | ----- | -------------------------------------- |
| `lMax` | int | Maximum spherical harmonic degree used |
| `energy` | float | X-ray energy in eV (e.g. `12500.0`) |
## Root datasets
| Path | dtype | Shape | Compression | Description |
| -------------- | ------- | ------ | -------------------- | ------------------------------------------------------- |
| `/q_grid` | float64 | `(Q,)` | ZFP lossless | Momentum transfer grid in angstrom^-1; `Q = len(qvals)` |
| `/sources_tsv` | uint8 | `(N,)` | Bitshuffle + Zstd-22 | Raw bytes of provenance TSV (optional) |
| `/makeup_tsv` | uint8 | `(M,)` | Bitshuffle + Zstd-22 | Raw bytes of ion makeup TSV (optional) |
Both TSV datasets are written once and never overwritten on subsequent runs.
## Molecule data -- `/<group>/<stem>/`
Each `.xyz` file produces one HDF5 group nested two levels deep.
```
/<group_name>/
<stem>.attrs['name'] str
<stem>/
I_q float32 (Q,)
coords float64 (n, 3)
angles float64 (n, 2)
r float64 (n,)
elms str (n,)
```
| Level | Key | Description |
| --------- | -------------- | ------------------------------------------------------------------------------------------------------------------- |
| group | `<group_name>` | Arbitrary label supplied via the `groups` dict argument |
| subgroup | `<stem>` | Filename without `.xyz` extension |
| attribute | `name` | Molecule name string (from XYZ line 2) |
| dataset | `I_q` | Orientationally-averaged scattering intensity, float32 `(Q,)`, ZFP lossless |
| dataset | `coords` | Centroid-subtracted Cartesian coordinates, float64 `(n, 3)`, ZFP lossless |
| dataset | `angles` | Spherical angles, float64 `(n, 2)`: col 0 = theta (polar, 0 to pi), col 1 = phi (azimuthal, 0 to 2pi), ZFP lossless |
| dataset | `r` | Radial distances from centroid in angstroms, float64 `(n,)`, ZFP lossless |
| dataset | `elms` | Element symbol per atom, variable-length UTF-8 string `(n,)`, uncompressed |
`Q` is the number of points in `/q_grid` and is fixed for the whole file. `n` varies per molecule.
Coordinates are centroid-subtracted (shifted to geometric centroid before storage). `angles` and `r` are stored pre-computed for fast loading; they are consistent with `coords` via:
```
r[i] = norm(coords[i])
theta[i] = arccos(z[i] / r[i]) (0 if r = 0)
phi[i] = arctan2(y[i], x[i])
```
Form factors are **not** stored -- they are recomputed from `xraydb`.
## Groups
The `groups` argument maps each group name to a directory of `.xyz` files. Every group becomes a top-level HDF5 group containing one subgroup per molecule.
| Group | Molecules | Atom range | Description |
| ----- | --------: | ---------- | ----------- |
| COD | 532,302 | 1-6,032 | Crystallography Open Database |
| QM9 | 133,844 | 3-29 | Small organic molecules |
| tmQM | 108,541 | 7-569 | Transition metal complexes |
| rcsb_sml | 96,158 | 28-6,036 | PDB small structures |
| viro3D | 60,488 | 173-6,046 | Viral protein structures |
| hydration_shells | 48,571 | 3-147 | Water solvation shells |
| rcsb_med | 31,749 | 2,996-6,046 | PDB medium structures |
| mofs | 30,863 | 10-5,760 | Metal-organic frameworks |
| (Na,Co,Ag,Pb,Mo,Fe)_monoatomic_clusters | 1,282 | 2-380 | Monoatomic clusters |
| binary_clusters | 371 | 2-1,482 | Binary alloy clusters |
| si_ge_clusters | 217 | 4-60 | Silicon/germanium clusters |
| ar_ne_clusters | 127 | 2-55 | Noble gas clusters |
| (NaCl)_nCl- | 70 | 3-71 | Sodium chloride clusters |
| **TOTAL** | **1,044,583** | **1-6,046** | |
## Compression codecs
| Codec | Used for | Notes |
| -------------------------------- | ---------------------------------------- | -------------------------------- |
| ZFP lossless (`reversible=True`) | `q_grid`, `I_q`, `coords`, `angles`, `r` | Floating-point; exact round-trip |
| Bitshuffle + Zstd level 22 | `sources_tsv`, `makeup_tsv` | uint8 blobs; ZFP incompatible |
`elms` is a variable-length UTF-8 string dataset and is stored uncompressed.
## B-tree corruption recovery (rcsb_med, June 2026)
The `rcsb_med` group B-tree was corrupted mid-build (at roughly 40% completion, ~40k of 101,989 entries written). Standard h5py operations on it (`del`, `keys()`) raised checksum errors. Recovery procedure:
### Step 1 -- OHDR binary scan
Scan the raw file with `mmap.find(b'OHDR')`, skip non-v2 headers (version byte != 2), then call `H5Oopen_by_addr` via ctypes on h5py's bundled libhdf5 to open each candidate object directly by byte offset, bypassing the corrupted B-tree. Each call is wrapped in a `signal.SIGALRM` timeout (1 s) to prevent infinite hangs on pathological corrupted objects. Valid molecule groups are written incrementally to a recovery file (checkpoint every 200 molecules for resume safety).
Result: 37 GB file, 4.7 M OHDR signatures, ~26 min, 25 timeouts.
**Warning -- zombie objects**: OHDR scan finds ALL HDF5 objects ever written to the file, including orphaned objects from previous build runs that were logically deleted but not physically zeroed. After recovery, cross-check every recovered key against the source XYZ directory and delete any key with no matching `<stem>.xyz`. In this run: 107,616 raw hits, 67,616 were garbage (old unprefixed hydration_shells orphans from a previous naming convention), leaving 40,000 legitimate rcsb_med entries.
### Step 2 -- Fresh file rebuild
`del hf['rcsb_med']` also fails with checksum errors on a corrupted group. Solution: build a new file from scratch using `h5py.File.copy()` (H5Ocopy -- raw chunk copy, no decompression) to transfer all intact top-level groups/datasets from the original, then copy rcsb_med from the recovery file. Rename rebuilt file over original.
Result: ~12 min to rebuild.
### Step 3 -- Resume build_db
With the recovered 40,000 entries in place, `build_db.py` resumes normally: it opens the file in append mode, skips entries that already exist, and fills in the remaining 61,989 rcsb_med entries plus all subsequent groups (rcsb_sml, si_ge_clusters, tmQM, viro3D).
### Key tools
- `h5clear -s <file>`: reset write-open flags left by an interrupted write
- `H5Oopen_by_addr` (ctypes): open HDF5 objects by raw byte offset, bypassing B-trees
- `signal.SIGALRM`: bound hanging C-library calls to a fixed timeout
## Crash safety
Entries are written under a temporary name `__tmp__<stem>` and atomically moved to `<stem>` only after shape assertions pass. Any `__tmp__*` keys found at startup are cleaned up before processing resumes.